WEBVTT

00:00:00.160 --> 00:00:06.640
how can this be faster than this when it should clearly be the other way around

00:00:04.960 --> 00:00:11.440
for years i've been reading comments from people who believe that the faster

00:00:08.559 --> 00:00:15.440
the gigahertz the faster the CPU and why shouldn't they believe that gigahertz

00:00:13.360 --> 00:00:20.240
also referred to as clock speed or frequency is quite literally a measure

00:00:18.240 --> 00:00:25.920
of how fast the transistors in a processor switch so all else being equal

00:00:23.359 --> 00:00:30.560
more gigahertz should be more better but all else is not equal and in today's

00:00:28.400 --> 00:00:34.960
video we're going to dive into what those unequal things are and just how

00:00:33.120 --> 00:00:38.800
unequal they can be we're also going to dive into today's sponsor arazi thanks

00:00:37.200 --> 00:00:42.800
to rozzy for sponsoring this video ahrazi's new oak yogurt webcams are

00:00:40.960 --> 00:00:46.879
privacy focused so you can be seen and heard only when you want to be get your

00:00:44.800 --> 00:00:50.239
okiya webcam with or without a ring light at the link down below

00:00:57.440 --> 00:01:04.799
to make sure that our test is as fair as possible both of our cpus used identical

00:01:02.000 --> 00:01:10.799
test benches ASUS tough b550 plus motherboards nocto nhd14 coolers 16 gigs

00:01:08.240 --> 00:01:16.720
of dual channel 3600 megahertz c14 memory a crucial p5 NVMe SSD and an rtx

00:01:14.400 --> 00:01:20.320
3060 xc from evga we're going to have all these parts in our affiliate links

00:01:18.159 --> 00:01:24.240
down below well most of them gpus can kind of be hard to find

00:01:21.600 --> 00:01:28.799
now for the cpus to keep politics out of the conversation we're going to be using

00:01:26.000 --> 00:01:33.840
only AMD branded processors but these principles can be applied to any other

00:01:31.040 --> 00:01:38.159
situation where cpus are being compared naturally we started with a full run of

00:01:35.840 --> 00:01:43.759
our benchmark suite at out of the box speeds so we can see how the higher

00:01:40.159 --> 00:01:45.439
gigahertz 3600 xt fared against the

00:01:43.759 --> 00:01:50.000
5600x remember that both of these cpus have

00:01:47.600 --> 00:01:54.240
exactly the same number of cores and threads

00:01:51.360 --> 00:01:59.360
somewhat intuitively the newer processor does outperform the older one and

00:01:56.640 --> 00:02:04.159
sometimes by a considerable margin but why well many modern processors are

00:02:02.159 --> 00:02:08.239
capable of dynamically boosting their clock speed under favorable conditions

00:02:06.320 --> 00:02:13.599
say for example when they have a really good cooler installed maybe our 5600x is

00:02:11.039 --> 00:02:18.640
just a mad CPU frequency boosting machine let's try raining it in and

00:02:15.760 --> 00:02:25.680
seeing what happens then at our locked clock speed of 3.4 gigahertz the 5600 xt

00:02:22.640 --> 00:02:28.000
still wins in every single test so

00:02:25.680 --> 00:02:33.440
clearly then gigahertz is not the only determining factor for CPU performance

00:02:30.800 --> 00:02:38.400
but these numbers aren't enough to tell the whole story let's look at gaming if

00:02:35.920 --> 00:02:43.840
i only measured average FPS in shadow of the tomb raider and grand theft auto 5 i

00:02:40.959 --> 00:02:49.040
might think that a 5600x is only about 5 percent faster than a 3600 xt in the

00:02:46.640 --> 00:02:54.239
real world but take something more CPU bound like cs go and these two cpus with

00:02:52.239 --> 00:02:58.160
the same core counts running at the same frequencies are nowhere near each other

00:02:57.200 --> 00:03:03.920
but then dropping the clock frequency even

00:03:00.319 --> 00:03:07.040
further to 2.4 gigahertz it's clear that

00:03:03.920 --> 00:03:10.239
the lower the clock goes the slower our

00:03:07.040 --> 00:03:12.400
cpus get so what is it does gigahertz

00:03:10.239 --> 00:03:18.159
matter or not there are a couple of takeaways here starting with that yes

00:03:15.200 --> 00:03:22.480
gigahertz absolutely matters which raises the question then why don't CPU

00:03:20.319 --> 00:03:27.360
manufacturers just run their chips at higher clock speeds i mean bring on the

00:03:24.400 --> 00:03:30.720
10 gigahertz cpus am i right well

00:03:28.239 --> 00:03:34.560
that was the plan actually but higher clock speeds come at the cost

00:03:32.400 --> 00:03:38.480
of more power consumption which tends to result in hotter running chips

00:03:36.959 --> 00:03:42.640
thankfully though there are a lot of other levers that CPU designers can pull

00:03:40.560 --> 00:03:48.480
to improve performance which leads us to our second takeaway cpus or any kind of

00:03:46.159 --> 00:03:53.840
processor for that matter gpus phone socs anything

00:03:50.560 --> 00:03:54.799
should never be compared using gigahertz

00:03:53.840 --> 00:03:59.200
alone it is clearly an important spec and

00:03:57.120 --> 00:04:03.280
manufacturers do need to disclose it because it enables us to compare

00:04:00.799 --> 00:04:08.879
products within their own families but if you want to talk about an m1 mac

00:04:05.680 --> 00:04:12.319
versus an Intel mac or an AMD GPU versus

00:04:08.879 --> 00:04:14.400
an NVIDIA one don't even bring it up you

00:04:12.319 --> 00:04:18.880
would only be revealing your ignorance on the subject let's talk then about

00:04:16.400 --> 00:04:22.880
some of the ways a CPU can differ aside from gigahertz an obvious one is that

00:04:20.959 --> 00:04:27.199
they can be designed to process more threads or tasks in parallel Intel was

00:04:25.440 --> 00:04:31.840
the first to process two concurrent threads on a consumer chip with hyper

00:04:29.040 --> 00:04:36.560
threading or smt while AMD was the first to build a truly multi-core CPU with

00:04:34.320 --> 00:04:41.040
their x2 series dual cores that were capable of doing nearly double the work

00:04:38.720 --> 00:04:45.840
under ideal conditions the only drawback to additional cores is that they

00:04:43.040 --> 00:04:50.880
increase die size meaning cost and power consumption and they can't be used to

00:04:48.320 --> 00:04:55.440
accelerate single threaded workloads so in many consumer applications like games

00:04:53.440 --> 00:04:59.840
they are only helpful up to a point currently AMD and Intel's mainstream

00:04:57.360 --> 00:05:04.400
lineups top out at 16 and 8 cores respectively so we can't keep pushing

00:05:02.320 --> 00:05:08.080
core counts forever and expect consumer applications to scale and clock speeds

00:05:06.560 --> 00:05:12.960
have been locked in the same range for over 15 years then what have they

00:05:11.120 --> 00:05:18.800
changed to really push forward single core performance the simple answer is

00:05:15.520 --> 00:05:21.919
ipc or instructions per clock if we

00:05:18.800 --> 00:05:23.759
think of a CPU like a mine and each core

00:05:21.919 --> 00:05:28.720
like a miner running back and forth doing work the clock speed is how many

00:05:26.800 --> 00:05:34.160
times our miner can run back and forth per second while the ipc is how much

00:05:31.360 --> 00:05:39.759
they can carry on each load look at the apple m1 for example joe average gamer

00:05:36.720 --> 00:05:41.440
might laugh at its meager 3.2 gigahertz

00:05:39.759 --> 00:05:46.919
clock speed but when it comes to the real world it performs pretty damn well

00:05:44.560 --> 00:05:53.039
like this sexy retro GPU t-shirt from lttstore.com what that tells us about it

00:05:49.600 --> 00:05:55.039
is that it has better ipc than a CPU

00:05:53.039 --> 00:06:00.240
that runs at a higher frequency but performs the same the problem though is

00:05:57.360 --> 00:06:03.120
ipc sounds a lot simpler than it is you can't just

00:06:01.440 --> 00:06:08.319
add more instructions to each clock cycle let's go back to our mine analogy

00:06:05.440 --> 00:06:13.520
the problem is that our mind contains every single possible type of mineral or

00:06:11.360 --> 00:06:19.360
rock and let's say those represent different apps or programs and each of

00:06:16.080 --> 00:06:20.960
them requires specialized equipment so

00:06:19.360 --> 00:06:24.639
let's say you level up your miner by adding more points to their shovel and

00:06:23.039 --> 00:06:29.199
suddenly there's a boost to your coal gathering but sifting for gold well

00:06:27.600 --> 00:06:33.039
shovel doesn't help you with that so performance is entirely unaffected

00:06:31.440 --> 00:06:38.560
that's how you can see a new generation of CPU come out that absolutely crushes

00:06:35.440 --> 00:06:42.080
cinebench but gets the same FPS in games

00:06:38.560 --> 00:06:44.240
so ipc is problematic along with clock

00:06:42.080 --> 00:06:48.639
speeds and core counts it's one of the most important ways to predict a

00:06:45.919 --> 00:06:52.639
processor's performance and yet unlike those other attributes

00:06:50.240 --> 00:06:55.919
nobody can agree on a fair and objective way to measure it the way that we

00:06:54.160 --> 00:07:01.360
enthusiasts use the term saying things like this new CPU has 20 higher ipc than

00:06:59.199 --> 00:07:05.360
the old one can be misleading a manufacturer could easily spend all

00:07:03.599 --> 00:07:09.520
their time tuning performance for a single commonly benchmarked program like

00:07:07.599 --> 00:07:14.080
geekbench or cinebench when that wouldn't be representative of the real

00:07:11.199 --> 00:07:18.240
world experience of using it though AMD and Intel also throw the term around in

00:07:16.160 --> 00:07:23.039
this way when it suits them so i blame them now there are major CPU

00:07:20.880 --> 00:07:28.240
design factors that can the real world performance of a high ipc CPU

00:07:26.000 --> 00:07:33.520
that's tuned for a particular benchmark let's talk about waste going back to our

00:07:30.639 --> 00:07:39.120
mine analogy adding cash to a CPU is kind of like making easy piles of our

00:07:36.319 --> 00:07:43.599
minerals or data that can be shoveled and carted out of the mine more quickly

00:07:41.520 --> 00:07:46.240
the bigger the pile the more likely it is that you can just fill up your

00:07:44.960 --> 00:07:50.639
wheelbarrow and off you go on the other hand if there's

00:07:48.560 --> 00:07:54.560
nothing in the pile the miner has to go deeper into the mine or to the system

00:07:52.720 --> 00:07:59.120
memory to retrieve it that's going to take longer then there's the branch

00:07:56.319 --> 00:08:04.080
predictor it is kind of like mine supervisors who attempt to proactively

00:08:01.440 --> 00:08:08.160
communicate which minerals are going to be needed in the near future rather than

00:08:06.240 --> 00:08:12.639
just having the miners wait around for an order CPU designers can dramatically

00:08:11.120 --> 00:08:17.039
improved performance with accurate branch prediction but the logic for it

00:08:15.120 --> 00:08:21.520
takes up space on the CPU that could also just be used to add more miners so

00:08:19.919 --> 00:08:26.000
it ends up being a delicate balancing act speaking of the physical layout of

00:08:23.599 --> 00:08:30.479
the course imagine if our miner parked their wheelbarrow right next to the

00:08:27.680 --> 00:08:34.959
mineral heap instead of five steps away and carried it like that CPU designers

00:08:32.959 --> 00:08:39.839
are always looking for ways to make each load more efficient and sometimes the

00:08:37.440 --> 00:08:45.120
actual physical proximity of CPU elements can be a big difference maker

00:08:42.560 --> 00:08:50.800
so an obvious solution to this problem then is to stop using gigahertz stop

00:08:47.920 --> 00:08:55.360
using ipc and rather use a broad industry standard set of real-world

00:08:53.120 --> 00:09:00.080
tests the problem with that is if we're looking at real-world benchmarks we end

00:08:57.360 --> 00:09:04.880
up with real-world messiness including politics between competing brands who

00:09:02.000 --> 00:09:09.600
would each naturally prefer tests that favor their own products this is why to

00:09:07.120 --> 00:09:13.680
this day we still need reviewers lots of them so that you can see a wide variety

00:09:11.680 --> 00:09:17.680
of different methodologies and test suites and how the product that you're

00:09:15.519 --> 00:09:22.080
considering stacks up and so that you can learn about sponsors like nordpass

00:09:20.080 --> 00:09:26.399
nordpass wants to help you keep your private information safe the nordpass

00:09:24.480 --> 00:09:30.880
password manager stores your passwords in a single place and recognizes your

00:09:28.720 --> 00:09:35.200
favorite website so it can automatically fill in your login details you can

00:09:32.480 --> 00:09:39.360
create new complex and secure passwords with the built-in password generator and

00:09:37.040 --> 00:09:44.480
then access those credentials on any device even when you're offline they

00:09:41.600 --> 00:09:48.800
offer unlimited password note and credit card storage and nordpass premium starts

00:09:46.480 --> 00:09:52.480
at just two dollars and fifty cents a month it comes with additional features

00:09:50.560 --> 00:09:57.040
like password health reports data breach alerts and up to six active devices for

00:09:55.040 --> 00:10:02.000
nord passes back to school sale for a limited time you can get 74 off a

00:10:00.000 --> 00:10:05.519
two-year nordpass premium plan with an extra four months for free so start

00:10:03.920 --> 00:10:09.760
protecting your passwords today at nordpass.com

00:10:07.120 --> 00:10:12.800
Linus and use code Linus as always thanks for listening folks i

00:10:11.519 --> 00:10:18.640
hope it helps you make the right choice next time you're looking to upgrade if you enjoyed this video hey give a thumbs

00:10:16.320 --> 00:10:25.399
up and make sure to check out is 4 core still enough you might be surprised by

00:10:20.800 --> 00:10:25.399
the results they are only helpful